EMSegmenter Tutorial (Advanced Mode)

Transcription

1 EMSegmenter Tutorial (Advanced Mode) Dominique Belhachemi Section of Biomedical Image Analysis Department of Radiology University of Pennsylvania 1/65

2 Overview The goal of this tutorial is to apply the EMSegmenter to MRI brain scans. We will segment the clinical T1 scan shown below into grey matter, white matter, and cerebrospinal fluid. The tutorial is based on. Before After 2/65

3 Overview We will segment the MRI scans by specifying a 'Task' for the EMSegmenter. The task captures the setting of the EMSegmenter for generating the automatic segmention of the subject scan. A task specifies the pre-processing of the scan, such as the type of atlas-to-image registration. It also specifies the structures to be segmented and the atlas specifying the structures. Furthermore, the task specifies the parameters related to the optimization algorithm (EM). Task Preprocessing Structures Atlas Parameters 3/65

4 Overview The tutorial leads you through the steps necessary for creating a new task: Step 1: Define task name and type of pre-processing Step 2: Define Input Channel Step 3: Define the Anatomical Tree Step 4: Assign an atlas to each node in the tree Step 5: Defining the Atlas to Image Registration Step 6: Further specify pre-processing Step 7: Specifying the Intensity Distribution Step 8: Define EM Specific Parameters Step 9: Specify the Region of Interest and complete the Segmentation 4/65

5 Define Task Step 1: Define task name and type of pre-processing The name should be a brief description of the segmentation scenario that the task addresses, such as 'T1 Brain Tissue Segmentation'. Each pre-processing type defines a sequence of approaches for modifying the scan before segmenting the scan into the structures of interest. For example, the pre-pocessing MRI Human Brain consists of image inhomogeneity correction and atlas registration. For further details please see 5/65

6 EMSegmenter (Advanced mode) Left-click on the Modules menu 6/65

7 Select EMSegmenter Module Select Segmentation EMSegmenter 7/65

8 Update Task List The first step of the EMSegmenter workflow appears. Click on Update task list to download from the web the current version of each task. 8/65

9 Create New Task Left-click on the Task menu and select Create new task 9/65

10 Create New Task Choose Tutorial as a new task name and select MRI Human Brain Pre-processing Click Apply Updated preprocessing tasks will contain a version number (e.g. MRI Human Brain V2) Please select the latest version. 10/65

11 Define Input Channel Step 2: Define Input Channel The EMSegmenter is equipped for multi-channel segmentations. For this tutorial, we want to perform single channel T1 segmentation. We now specify the task accordingly by loading in a T1 scan and creating a single input channel. 11/65

12 Define Input Channels Click OK 12/65

13 Load subject volume 2. To load the subject data click on the File menu select Add Volume 1. Download our MRI volume mentioned below and rename it to t1.nrrd 13/65

14 Load Subject Data Browse to your download location, select t1.nrrd, And click on Apply. 14/65

15 Define Input Channel Click on Add Channel. Type 'T1' into the Name field. Assign the Volume t1. Click Next 15/65

16 Define Input Channel To confirm click Yes 16/65

17 Define Anatomical Tree Step 3: Define the Anatomical Tree In this step we are defining the anatomical structures we want to segment and store the information in a tree data structure. Each node represents an anatomical structure. Additionally, a label and color can be assigned to each node, which are used when generating the segmentation map. 17/65

18 Define Anatomical Tree Right-click on Root, and select Add sub-class 18/65

19 Define Anatomical Tree A new node appears 19/65

20 Define Anatomical Tree Change the node name to Background Known KWWidgets Bug! This field cannot be empty. Sorry! 20/65

21 Define Anatomical Tree Right-click on Root, and select Add sub-class 21/65

22 Define Anatomical Tree Rename the second sub-class Intracranial Cavity The anatomical tree contains two components: Background and Intracranial Cavity 22/65

23 Define Anatomical Tree Right click on Background, and select Add sub-class. Add the sub-class Air to Background. Click on the colored box next to Color A new widget appears, choose Default Labels from file Slicer3_2010_Brain_Labels Set the label value for Air to 0. 23/65

24 Define Anatomical Tree Add the Skull sub-class to Background, and set the label value for Skull to 0. 24/65

25 Define Anatomical Tree Using the same process, right-click on Intracranial Cavity, and add the three following Sub-classes: - Grey Matter, label 9 - White Matter, label 10 - CSF, label 5 Click on Next to assign the atlas to the structures 25/65

26 Define Atlas Step 4: Assign an atlas to each node in the tree We now further characterize each anatomical structure by specifying the atlas associated with that structure. For the EMSegmenter, the atlas defines the spatial distribution of the structure of interest, which is the frequency the structure appeared at each image location in a given set of scans. For further information on generating these atlas please read: L. Zöllei, M. Shenton, W.M. Wells III, K.M. Pohl. The Impact of Atlas Formation Methods on Atlas-Guided Brain Segmentation, Statistical Registration. In Pair-wise and Group-wise Alignment and Atlas Formation Workshop at MICCAI 2007: Tenth International Conference on Medical Image Computing and Computer-Assisted Intervention, pp , /65

27 Define Atlas In the following steps we are assigning atlas volume data to each structure. In this tutorial you only have to touch with Select Probability Map The field Select Parcellation Map can be ignored. 27/65

28 Load Atlas Data Click on File Add Data... 28/65

29 Load Atlas Data Click on Add File(s) Browse to your Slicer3 install directory and from there to./share/slicer3/modules/emsegment/tasks/mri-human-brain/ Select the six atlas data files and click Open 29/65

30 Load Atlas Data Click on Label None to uncheck all LabelMap checkboxs Click Apply 30/65

31 Load Atlas Data The loaded atlas data appear in the viewer. 31/65

32 Define Atlas Select Air in the anatomical tree. Left-click on Select Probability Map and assign the probabilistic atlas atlas_air to the Air structure. 32/65

33 Define Atlas Select Skull in the anatomical tree. Left-click on Select Probability Map and assign the probabilistic atlas atlas_skullneck to the Skull structure. 33/65

34 Define Atlas Select Grey Matter in the anatomical tree. Left-click on Select Probability Map and assign the probabilistic atlas atlas_greymatter to the Grey Matter structure. 34/65

35 Define Atlas Select White Matter in the anatomical tree. Left-click on Select Probability Map and assign the probabilistic atlas atlas_whitematter to the White Matter structure. 35/65

36 Define Atlas Select CSF in the anatomical tree. Left-click on Select Probability Map and assign the probabilistic atlas atlas_csf to the CSF structure. Click on Next 36/65

37 Edit Registration Parameters Step 5: Defining the Atlas to Image Registration In general, the currently defined atlas has to be aligned to the subject scan. To do so, we define in this step the template, which in this case is a T1 scan, that the atlas is currently aligned to as well as the type of registration we would like to perform 37/65

38 Edit Registration Parameters Select atlas_t1 to assign the atlas to the input channel T1 38/65

39 Edit Registration Parameters Select Fast for the Affine Registration and the Deformable Registration. Set Package to BRAINS Click on Next 39/65

40 Define Preprocessing Step 6: Further Specify Preprocessing In the first step, we defined the type of preprocessing we wanted to perform. We now further specify the pre-processing by answering a set of questions further specifying the type of data we attend to segment. For example, in this tutorial we assume that the subject scan is already aligned to the atlas so that we skip the atlas to image registration during preprocessing. 40/65

41 Define Preprocessing We note, that in this tutorial the subject data set is image inhomogeneity corrected and pre-registered to the atlas. Thus, the 'registration flag' and the 'inhomogeneity correction flag' are not checked. Please do not check for this tutorial as pre-processing can be time consuming. Click on Next 41/65

42 Define Preprocessing The EMSegmenter will perform some standard pre-processing. Click on Yes to confirm. 42/65

43 Define Preprocessing Please wait until pre-processing has been finished 43/65

44 Specify Intensity Distribution Step 7: Specifying the Intensity Distribution In this step, users further specify each anatomical structure by defining the intensity distribution that is typical for the structure in the input scan. In this tutorial the step can be skipped as the intensity distributions have been calculated during the pre-processing. 44/65

45 Specify Intensity Distribution Click on Next 45/65

46 Edit Node-based Parameters Step 8: Define EM Specific Parameters The EMSegmenter segments the input scans of Step 1 into the structure of interest of Step 2 by using an optimization algorithm called the Expectation Maximization Algorithm. This algorithm has specific parameters that influence the segmentation. In this tutorial we will specify: - class weights, which define the relative importance of structure over other structure. This is useful if a structure is too dominant in the automatic segmentation. By lowering the weight, the structure will be less present in the corresponding automatic segmentation. - atlas weight, which define the importance of the atlas (of Step 3) over the image data defined in Step 1. One might want to lower the weight if the intensity distributions clearly define each structure to be segmented. - Input Channel weight, which defines the importance between the different input channels for the structure of interest. Since we only defined one input channel, this parameter should simply be set to 1. - Alpha, which defines the smoothness of the segmentation. The alpha value has to be between 0 and 1. An alpha value of 1 produces fairly smooth segmentations while an alpha value of 0 generally results in noisy segmentations. 46/65

47 Edit Node-based Parameters Per default all the EM Input Parameters are unspecified. 47/65

48 Edit Node-based Parameters Left click on Background and Enter the following parameters: -Class Weight Atlas Weight 1 We only defined one input channel, please set Input Channel Weights: T1: /65

49 Edit Node-based Parameters Left click on Intracranial Cavity And enter the following parameters: -Class Weight Atlas Weight 1 -Input Channel Weights: T1: /65

50 Edit Node-based Parameters Enter the following parameters for Air and Skull Air: Class Weight: 0.7 Atlas Weight: 1.0 Input Channel Weight: 1.0 Skull: Class Weight: 0.3 Atlas Weight: 1.0 Input Channel Weight: /65

51 Edit Node-based Parameters Enter the following parameters for GM, WM, and CSF GM: Class Weight: 0.45 Atlas Weight: 0.01 Input Channel Weight: 1.0 WM: Class Weight: 0.3 Atlas Weight: 0.7 Input Channel Weight: 1.0 CSF: Class Weight: 0.25 Atlas Weight: 0.01 Input Channel Weight: 1.0 Click on Next 51/65

52 Run Segmentation Step 9: Specify the Region of Interest and complete the Segmentation This is the last step of the EMSegmenter wizard. The Volume Of Interest (VOI) can be specified, and one can start the EM algorithm, which will segment the input channels by taking all the information entered in the previous steps into account. 52/65

53 Run Segmentation In the first run we don't specify a volume of Interest (VOI) Click on Segment 53/65

54 Run Segmentation The EM algorithm is running Please wait The processing time on a standard computer is about 2 minutes 54/65

55 Results: Run Segmentation White matter The results of the EM Segmentation are overlaid on the T1 volume. Grey matter CSF 55/65

56 Consecutive adjustment As previously mentioned, one might want to adjust the parameters of Step 8 in order to improve the segmentation. We now adjust three parameters and show the impact on the segmentation The following slides illustrate -how to specify a volume of interest and -how to adjust segmentation parameters that refine the segmentation result. 56/65

57 Volume Of Interest (VOI) To specify a smaller volume of interest, make it first visible by selecting the checkbox Display VOI in 2D Viewer, adjust the size of the VOI by moving the 'Range' slider, unselect the checkbox Display VOI in 2D Viewer, and click Segment. 57/65

58 Result: Volume Of Interest (VOI) Only the VOI has been segmented. Note that a smaller VOI leads to a faster segmentation. For the next adjustment click on Back 58/65

59 Adjusting Parameters Step 8/9. Edit Node-based Parameters: We want to change the class weight for grey matter and automatically update the class weight for white matter. To do so, select the checkbox next to white matter and change the class weight for grey matter to Click on Segment. 59/65

60 Result: Adjusting Parameters The result of the new segmentation based on the changed parameters appears. This process can be continued to get a better segmentation. 60/65

61 Compare Results We are changing the EM algorithm parameters in the following 3 slides. Please overlay the different results in the GUI. (e.g. EM_Map / EM_Map1) Use the slider to fade between background and foreground layers. 61/65

62 Low ICC alpha value Effect: The labelmap Is less smooth 62/65

63 Low white matter atlas weight Effect: Finer white matter structures become visible 63/65

64 High grey matter class weight Effect: Overestimation of grey matter 64/65

65 Further Info & Acknowledgments EMSegmenter Wiki Page: The EMSegmenter technology behind was reported in: K.M. Pohl et. A hierarchical algorithm for MR brain image parcellation. IEEE Transactions on Medical Imaging, 26(9), pp , We thank the following institutions for their support: 65/65